Financial-risk aware machine learning for industrial anomaly detection

Digital transformation in asset-intensive industries is constrained by the separation between Information Technology (IT), which supports enterprise decision-making, and Operational Technology (OT), which monitors and controls physical assets. This study proposes a Financial-Risk Aware Digital Twin framework that combines machine-learning-based anomaly detection with event-specific economic consequence analysis. Using the benchmark 3W oil and gas dataset, an XGBoost classifier was developed to identify operational states including severe slugging, spurious Downhole Safety Valve closure, flow instability, productivity loss, scaling, and hydrate formation. The model achieved 93% overall accuracy, with weighted-average precision and F1-score of 0.94 and 0.93, respectively. Predicted class probabilities were then converted into class-specific expected financial exposure using illustrative consequence scenarios. At a predicted probability of 20%, the corresponding exposure was $200,000 for hydrate formation, $100,000 for DHSV closure, and $10,000–$20,000 for one day of severe slugging under the selected scenario assumptions. Financial weighting also changed event prioritization in an illustrative re-ranking example. For hydrate formation, a base-case failure consequence of $1,000,000 and preventive-action cost of $5,000 produced a 0.5% break-even intervention probability; sensitivity analysis across alternative assumptions yielded thresholds from 0.125% to 2.0%. The results show how probabilistic condition monitoring can be extended into an economically informed decision-support framework that prioritizes operational responses according to both likelihood and consequence.